Big difference in the evaluation between Stockfish 10 and Leela Chess Zero

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
Javier Ros
Posts: 142
Joined: Fri Oct 12, 2012 10:48 am
Location: Seville (SPAIN)
Full name: Javier Ros

Big difference in the evaluation between Stockfish 10 and Leela Chess Zero

Post by Javier Ros » Tue Jan 01, 2019 8:13 pm

A little over a year ago, nobody in this forum would have thought that any program or human would frequently defeat Stockfish positionally over the board. Its analysis power leaves no doubt of its first place among the computer programs.
But the appearance of Alpha Zero and Leela Chess Zero has changed things.

In the next game, the great disparity in the evaluation between Stockfish 10 and Leela Chess Zero shows the difference between the approaches of both.
Lc0 32374 played on RTX2070 and Sf10 on i7 4790S 6 cores, with 5 minutes plus 1 second per move.

This is the second position of the Balsa 25 Top Set, with thousands of games played between the best human players



the initial evaluations show the equality of the the game: 6. e4 {0.35/13 6} Bb4 {0.36/27 20} 7. Nxc6 {0.31/13 3}

then 7... dxc6 {0.29/28 5} by Sf and Lc0 thinks to have some advantage 8. Qxd8+ {0.80/14 5}

after 13.h5 {0.84/11 13} Bb7 {0.00/26 5} 13... h6 was better to avoid 14.h6



while Sf assesses the position as equaled to slight disadvantage, Lc0 thinks that its advantage is increasing

14. h6 {1.05/10 8} g6 {0.00/29 2} 15. f3 {1.20/10 7} Rag8 {0.00/30 12} 16. Rd1 {1.56/10 9} Re8 {0.00/30 11} 17. Bd3 {1.78/11 8} Rhf8 {-0.01/27 4} 18. O-O {1.85/12 7} Ba6 {-0.43/29 9} 19. Rfe1 {2.26/13 6} Nc5 {-0.39/30 4}

after the change of rooks the control of the black squares by the good bishop is evident while the black bishop is bad

23. Red1 {2.67/15 5} Bc8 {-0.48/30 2} 24. Rxd7+ {4.06/14 8} Bxd7 {-0.91/28 9} 25. Be7 {4.25/13 3}



the sacrifice of the c3 pawn after 25..Rg8 {-1.17/31 21} 26. Kf2 {5.45/12 8} is the price to get a winning position that Lc0 takes more than 100 moves to turn into victory



Repeating the game at 15 + 1, we see that the positional style of Lc0 does not always work even with exchange sacrifice 27. Rxd7+ {+4.92/19 9} Kxd7 {-1.46/38 25}. In the disparity of evaluations this time Sf10 was right.




The love relationship between a chess engine tester and his computer can be summarized in one sentence:
Until heat do us part.

Post Reply